Research And Applications Of Imbalanced Image Classification Problem Based On Deep Learning

As one of the hot research directions in computer vision,image classification has become an essential foundation for pattern recognition,artificial intelligence,and other related areas.The emergence of deep learning technology promotes the development of image research.With the continuous breakthrough of deep learning in computer vision,deep neural networks have greatly improved various evaluation indicators of image classification with their powerful learning ability.However,different image data sets have different problems among the massive data,limiting the recognition ability.In addition,the imbalanced problems in image classification tasks are widespread in real-world applications.They have become one of the major problems in machine learning and data mining in recent years.High-dimensional images' highly imbalanced classification problem still faces significant challenges,and the difficulty will be greatly increased,especially when the minority samples are minimal.At present,the samples generated by the over-sampling-based methods are easily confined in the vicinity of a limited number of real samples,resulting in the balance degree of data distribution being far lower than that of the number of samples.Therefore,research on highly imbalanced and high-dimensional image classification problems with rare minority samples is of high academic significance.The qualitydiversity trade-off for the generated samples is one of the most crucial objectives in imbalanced problems.The existing methods mainly include sample-level operations,model-based searches,or increasing diversity by adding noise,but feature-level research is lacking.Therefore,the techniques at the feature level need to be further explored.By analyzing the characteristics and demands of the problems mentioned above,this doctoral thesis will incorporate related biological visual mechanisms and evolutionary algorithms into deep learning technologies.The main research work is as follows:(1)Considering the easily confused categories and the obscure objectives against the background patterns in image classification tasks,we propose a visual long-short-term memory(VLSTM)based neural network integration model inspired by the visual perception process and the visual memory mechanism in the biological vision system.The VLSTM model consists of a visual perception module,a visual short-term memory module,and a visual long-term memory module parallelly.According to the characteristics of different modules,we select and design the corresponding network models to simulate the biological vision system to extract the visual perception information,visual short-term memory information,and visual long-term memory information of pictures.The VLSTM integration model is evaluated on two public data sets.The experimental results verify that the features extracted from different modules are complementary to a certain extent and accord with the characteristics of the visual information represented by their respective modules,showing that the VLSTM achieves superior classification performance compared with the comparison algorithms.(2)For the highly imbalanced and high-dimensional image classification problem with a limited number of minority samples,to realize the knowledge transfer from the majority samples to the minority samples,we propose a conditional variational auto-encoders(CVAEs)based SelfTransferred(CVAE?Se Tred)algorithm to mine the correlation and the distribution similarity between the minority samples and the majority samples in the same data set.The CVAE?Se Tred model integrates distributed learning,self-transferred learning,image generation,and dataset rebalancing in a unified architecture.We verify the effectiveness of the CVAE?Se Tred model on two imbalanced data sets,and the experimental results show that the proposed model is stable and convergent.Finally,our model can learn domain-invariant and multivariate Gaussian distributed latent features.(3)Since the proximity between the generated data distribution and the real data distribution determines the quality and the diversity of the generated data,to address the trade-off problem between the quality and the diversity in imbalanced learning,we propose a Modified Estimation Distribution Algorithm based Latent feat Ure Distribution Evolution(MEDA?LUDE)algorithm to learn the optimal distribution of latent feature space through combining deep neural networks and evolutionary algorithms.The proposed model explores the effect of the Large-margin Gaussian Mixture(L-GM)loss function on distribution learning under the more complex conditions we assume.To meet the trade-off between quality and diversity,we also design a specialized fitness function based on the similarity among samples and modify the estimation of distribution algorithm to evolve the latent feature distribution,guiding the search for the latent feature space distribution.Finally,the MEDA?LUDE algorithm can generate samples with both the quality and the diversity on two imbalanced data sets constructed from the benchmark data sets.(4)We provide specific solutions to the fabric defect classification problems and its imbalanced problems under the textile industry background.Firstly,relying on the hardware acquisition device provided by Engineering Research Center of Digitized Textile ? Apparel Technology,Ministry of Education,Donghua University,we collect,process,and construct the Donghua University fabric defect data sets as well as the imbalanced data set.In addition,based on the original fabric defect images of Alibaba Cloud,which are publicly available,we also construct the Alibaba Cloud fabric defect data set and the imbalanced data set.Since fabric defect data sets exist the problems that confused categories are hard to distinguish,certain defects are obscure against the texture background,and data is imbalanced,we research on the corresponding deep network models and design specific network parameters according to the structural properties and the data characteristics of fabric defect images.Finally,the experimental results on the data sets we construct show that the network models can effectively finish the fabric defect classification tasks compared to other methods.